Academics rejoice!!
Less than a week after Anthropic's Claude Science launched, the open-source community has presented its own response.
An AI research team incubated by YC has released OpenScience, an open-source alternative to Claude Science.
It is an all-in-one AI research platform covering the entire workflow—from literature review and hypothesis generation to code experimentation and paper writing—but it is not tied to any single model provider.
DeepSeek, GLM, Claude, GPT… whether domestic or international, use whichever you prefer.
Moreover, the project uses the developer-friendly Apache 2.0 license and can be installed with just one command.
As soon as the news broke, the project shot straight to the top of X's trending list. People openly said:
This is what scientific AI should look like. (Company A: Just mention my name.)
Claude Science is powerful, but I can't use it...
Approximately five days ago, Anthropic officially launched Claude Science at a closed-door event hosted by MIT Technology Review.
This is an AI work platform designed specifically for scientists, offering a variety of tools and software packages commonly used by researchers.
For example, previously, a researcher had to search for papers on PubMed, write code in Jupyter, run statistical analyses in R, connect to a cluster via SSH to submit jobs, and then use various tools to create figures and write papers.
Switching between dozens of windows alone drains a huge amount of energy just through context switching.
And what Claude Science wants to do is fit all of this into a single workspace.
Specifically, it has made several key integrations:
At the database and toolchain level, it includes over 60 scientific database connectors and preconfigured skill packages covering common research areas such as genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.
Ask in natural language—the professional agent will automatically query across databases such as UniProt, PDB, Ensembl, ChEMBL, and GEO without requiring you to check each one individually.
It also integrates with NVIDIA’s BioNeMo Agent Toolkit, enabling direct connectivity to life science models such as Evo 2, Boltz-2, and OpenFold3.
At the implementation level, it introduces a multi-agent architecture.
The main agent handles overall planning, sub-agents process different tasks in parallel, and a reviewer agent is dedicated to fact-checking, such as verifying citations, validating calculation results, and flagging potential errors.
The generated results are not limited to text output—it can natively render 3D protein structures, genome browser tracks, and chemical structures.
And each chart will simultaneously retain the generated code, runtime environment, natural language explanations, and full conversation history.
In certain scenarios, scientists can directly modify a diagram with a single sentence, and the system automatically rewrites the underlying code.
At the computing power level, Claude Science can directly integrate with your existing lab infrastructure.
You can use laptops, Linux servers, or HPC cluster login nodes—connect via SSH or use a Modal account to on-demand access cloud GPUs, scaling from a single GPU to hundreds of GPUs.
Large datasets need to be loaded only once; sensitive data never leaves your own system—only the context required for each analysis step is sent to Claude.
Early beta users have already generated some real-world cases.
Neuroscientist Jérôme Lecoq from the Allen Institute used it to build a multi-agent "computational review template" with about 20 custom skills, enabling sub-agents to read thousands of papers, extract key insights and quantitative data, and generate reviews chapter by chapter.
Just put it this way: previously, writing a review took two years, but now he already has about ten on hand—
Many exceed 100 pages, and all citations have been verified by the Reviewer Agent.
Stephen Francis from the UCSF Brain Tumor Center uses it for molecular epidemiology studies of gliomas, running germline variant analyses.
He said Claude Science reduced the original time required to one-tenth, and his team independently verified the results, confirming that the analysis is both fast and reliable.
In light of Harvard physicist Matthew Schwartz’s evaluation of AI’s research capabilities in March this year, Claude’s current level is approximately equivalent to that of a second-year graduate student.
He published a guest article titled "Vibe Physics: The AI Grad Student" on Anthropic's official blog, documenting his entire process of using Claude Opus 4.5 to complete a theoretical physics paper.
At that time, his conclusion was:
The current scientific research capability of AI is roughly equivalent to that of a second-year graduate student: it can get work done without complaint, but requires constant supervision at every step.
This judgment was later incorporated by Anthropic into Claude Science’s technical documentation as a reference point for product positioning.
However, Claude Science currently has several hard limitations:
Supports macOS and Linux only
Available exclusively to Pro/Max/Team/Enterprise paying users
Only Claude's own models can be used on the platform.
When these barriers are stacked together, especially for domestic research teams, Claude Science becomes something unattainable.
Good news: An open-source alternative has arrived
In response to the above limitations, the open-source project OpenScience has emerged.
The team behind it is called Synthetic Sciences, founded in San Francisco in 2025, and graduated from the YC 2026 Winter batch this year.
The founding team has ambitious goals: to build a platform where scientists can directly delegate complex research tasks to AI co-scientists, enabling end-to-end automation from literature reviews and hypothesis generation to experiment execution and paper writing.
They have an internal core judgment:
Scientific foundation models must possess genuine "research taste," which cannot be achieved simply by scaling up parameters; instead, they require a dual approach—integrating product development with model training—to collect high-quality research process data through products and then use that data to train models with refined taste.
OpenScience is the first product to be realized along this path.
Although OpenScience shares the same mission as Claude Science, there is a fundamental difference between them:
Model-agnostic.
In the words of Synthetic Sciences:
Scientific AI should be open; no single company should monopolize the tools humanity uses for exploration and discovery, nor should it decide who is eligible to use them.
On this platform, you can directly connect Anthropic, OpenAI, Google, DeepSeek, GLM, and more—as long as you have an API key.
You can even run local models using Ollama, with not a single byte of data leaving your machine.
Your key remains on your local device, and requests are made directly to the model provider without passing through any intermediate servers.
Moreover, OpenScience supports model slicing on request.
Within the same workspace, you can use Claude for one step and switch to DeepSeek for the next, without changing any settings.
On a functional level, OpenScience is even more radical than Claude Science—
Includes over 250 built-in research skill packs—more than four times that of Claude Science—covering areas such as machine learning, computational biology, and cheminformatics, all of which are readable, editable, and extensible.
Installation is also simple—just run one command in the terminal:
Ready to use—your workspace automatically pops up in the browser. On first launch, select a model source and enter your API key to get started.
You can also install globally:
If managing a key is inconvenient, the team also offers a custodial platform called Atlas—
Deposit funds directly using multiple cutting-edge models without needing to configure keys individually, plus enjoy persistent research graphs and cloud-based computing power.
However, Atlas is not mandatory—you can run OpenScience with your own key and use it completely free of charge, with no barriers.
One More Thing
Interestingly, at the very bottom of the OpenScience GitHub page, you’ll find a specially added disclaimer:
OpenScience is an independent project. It is not affiliated with, endorsed by, or sponsored by Anthropic. “Claude” is a trademark of Anthropic, PBC, used here solely to describe compatibility.
We are an independent project and have no relation to Anthropic. Mentioning "Claude" is solely regarding compatibility—please don't overthink it.
It seems the initial "lobster" left a lasting impression on the entire open-source community.
OpenClaw changed its name several times before; this time, OpenScience has permanently embedded the statement distancing itself in the first version of the README.
Nothing else—first, survive, then talk about alternatives (doge).
Open source address:
https://x.com/i/trending/2073904804829741364?s=20
Reference link:
[1]https://x.com/SynScience/status/2073829478393086311?s=20
[2]https://x.com/i/trending/2073904804829741364?s=20
[3]https://www.openscience.sh/
[4] https://www.anthropic.com/news/claude-science-ai-workbench
This article is from the WeChat public account "Quantum Bit," authored by Yi Shui.